Chapter 16 · AI in science and the physical world2020
AI for science: protein folding
AI models predict a protein's 3D shape from its sequence, a problem that took scientists decades per protein.
Chapter 16 · AI in science and the physical world2020
AI models predict a protein's 3D shape from its sequence, a problem that took scientists decades per protein.
H: water-hating beadP: water-loving beadH–H contact
0 contacts now4 in the best fold
Ten beads on a flat grid. Beads of the same water-hating kind (pink H) want to huddle together, away from the water. The chain curls until as many H beads as possible touch.
Your body makes thousands of proteins. Each is a long chain made from 20 kinds of small building blocks called amino acids. The order of the blocks is the protein’s sequence, and the cell’s instructions (genes) spell it out.
Each joint can go straight (↑), turn left (↰) or turn right (↱). Try to get as many pink H–H contacts as you can without two beads landing on the same spot.
H: water-hating beadP: water-loving beadH–H contact
0 your contacts4 best possible4,067 allowed shapes
Keep bending. Look for H beads that could sit side by side.
This toy has only 8 joints, so the computer can check all of them in a blink. A real protein has hundreds of blocks and each can twist many ways, so the number of shapes explodes. That is why a model that predicts the answer is such a big help. This is a simplified teaching model, not how real prediction works.
What decides what a protein does?
Why can’t a computer simply try every possible fold of a real protein?
How should scientists treat a predicted structure?